Node.js CLI · Local browser control panel · Claude Code · Codex CLI · Private GitHub alpha
Agent Bridge
One terminal room for Claude, Codex, and project memory.
A local-first continuity layer for AI-assisted development. Agent Bridge is one terminal room where Claude and Codex are workers inside the session, while the tool owns project memory, decisions, handoffs, context pressure, and browser-based routing. The repo is private during alpha; GitHub access is available by request.
- Node.js
- JavaScript
- Local CLI
- Claude Code CLI
- Codex CLI
- Vanilla HTML/CSS/JS
- Local HTTP server
- File-backed memory
- Tailscale-ready access
- GitHub private alpha
Origin
How it started
The pain was not that AI tools could not answer questions. The pain was that every new session forgot the project: what was decided, why it mattered, what changed, and what future-me should know before touching the code again. Agent Bridge is the answer to that gap — not another remote terminal, but the memory and decision layer that keeps AI-assisted work continuous.
Features
What it does
One terminal room
@claude, @codex, @both, and @debate route work to installed local CLIs from one project-aware prompt. Claude can handle product thinking and critique; Codex can handle implementation planning, debugging, and code review; Agent Bridge keeps the shared context.
Durable local project memory
/remember, /preference, /decide, /checkpoint, /compact, /handover, /timeline, and /why write durable continuity files into .agent-bridge/ so project state survives compaction, tool switching, and future sessions.
Local browser control panel
agent-bridge serve opens a local dashboard with diagnostics, current state, recent decisions, pinned notes, memory, preferences, timeline items, context summary, search, handoff actions, and browser-based agent routing.
Context packets for other AI tools
/packet and the browser's context-packet flow select relevant saved memory and format a copy-ready packet for ChatGPT, Gemini, Perplexity, Aider, or any other AI without pasting the whole project again.
Context pressure and health checks
agent-bridge doctor, /limits, /context, and /performance make context size, truncation, installed tool state, git state, and speed/depth choices visible instead of hidden in a stalled AI session.
Private remote access path
Designed around local-first safety. Remote access is framed through private Tailscale Serve inside the user's tailnet, not a public internet control panel that can call local tools without a real auth and audit model.
Under the hood
Engineering
Local-first architecture, by design
Agent Bridge has three layers: a Node.js CLI that owns memory and routing, a local website served from 127.0.0.1, and a browser control panel that talks only to the local app. The public website can market and document the product later, but the engine has to run on the user's machine because it reads project files and calls local Claude/Codex binaries.
Continuity as a file-backed system
Project state lives in .agent-bridge/: memory, preferences, decisions, pins, context summaries, current state, timeline, stack, structure, sessions, and runs. The product treats those files as the working memory of the project. Commands do not just answer; they leave a trail future sessions can use.
Routing, synthesis, and saved decisions
The first proof loop was Agent Bridge reviewing itself. A @both prompt routed the same question to Claude and Codex with the saved product context, then synthesized the answers into a concrete verdict and naming decision. The important part was not asking two AIs the same thing; it was saved memory changing the quality of the answer.
Browser dashboard with local write surfaces
The local dashboard exposes API surfaces for dashboard state, status, memory, context, search, saving continuity, and routing prompts. It can save memory, decisions, preferences, and pinned notes; trigger handover, compact, and timeline actions; and route questions to Claude, Codex, both, or a structured debate.